Raw sequence modeling
Training models closer to the original data representation, with fewer brittle assumptions in the input pipeline.
Early-stage AI infrastructure startup
untokenized
We are building tokenization-free language models that learn directly from raw sequences.
Our work focuses on byte-native model infrastructure for AI systems that can adapt across languages, code, and specialized domains without fixed handcrafted tokenizers.
untokenized is an early-stage AI startup developing core research and infrastructure for the next generation of language models, where segmentation is learned as part of the model rather than imposed as a preprocessing step.
Focus
Tokenization-free AI systems
Adaptive compression
Learning compact internal representations that can vary with content, context, and downstream predictive needs.
Infrastructure-first research
Building practical training, evaluation, and scaling systems for byte-native model experiments.
Platform
Designed for cloud-scale experimentation
Our near-term roadmap depends on accelerated training runs, controlled ablations, and repeatable evaluation infrastructure. Cloud credits help us move from prototype research to reliable model development workflows.
Compute
GPU-backed training and evaluation
Data
Raw byte and mixed-domain sequence pipelines
Research
Reproducible experiments and benchmark tracking
Status
Private prototype, startup trajectory
We are keeping implementation details private while developing the first internal prototypes. The company direction is simple: remove static tokenization as a hard dependency for capable language models.
Contact
Work with untokenized
For cloud credits, compute partnerships, or technical conversations, reach us by email.
General inquiries
segu@untokenized.segundosantos.com